The models we build ourselves.
Theo orchestrates the best engines from every lab and shows which one answered. These are the ones we make and run on our own hardware: 4 engines across 3 kinds of work, 1 live in Theo today and 3 rolling out. Every page says what the engine does, what it guarantees, what we measured and when, and where it is available.
Measured on our own hardware, dated, never a benchmark claim · free, no sign-in · llms.txt · JSON
The roster
4 engines, each with its own page
Live first, then the ones rolling out. Every card's numbers are our own measurements, dated; open a page for how the engine works, its safety controls and where it runs.
Theo Moonlight 1.1
Image engine · version 1.1
High fidelity image renders on our own hardware. Renders about one megapixel from a written prompt, at seven aspect ratios from square to 16:9.
- Render time
- 33 seconds
- A 1024 by 1024 image at the default quality setting. Measured Sep 15, 2026.
- Output
- About 1 megapixel
- Seven aspect ratios from 1:1 to 16:9, each around one million pixels. Measured Sep 15, 2026.
Read about Theo Moonlight
Theo Real World 1.1
Capture engine · version 1.1
Photos of a real thing become a 3D capture of the thing in its space. Takes up to 60 photos or one walk-around video of one subject and returns a capture you can orbit, walk through and view in AR.
- Quick capture
- 127 seconds
- From submit, on a 9 second walk-around clip; 40 camera positions solved from 120 candidate frames. Measured Sep 21, 2026.
- Full capture
- 269 seconds
- The same clip, end to end; a 10.5 MB capture that reproduces held-out frames faithfully. Measured Sep 21, 2026.
Read about Theo Real World
TheoArca Magnus 5.1
Text engine · version 5.1
Theo's own deep-reasoning engine, governed and audited. Reasons deeply by default, with a 1M-token context window and answers of up to 131K tokens.
- Context window
- 1,000,000 tokens
- The whole conversation, documents included, in one request. Measured Sep 3, 2026.
- Longest answer
- 131,072 tokens
- The gateway's output ceiling per request. Measured Sep 3, 2026.
Read about TheoArca Magnus
TheoArca Velox 5.1
Text engine · version 5.1
Theo's own low-latency engine, governed and audited. Answers fast, with the same 1M-token context window and the same 131K-token answer ceiling as Magnus.
- Context window
- 1,000,000 tokens
- The whole conversation, documents included, in one request. Measured Sep 3, 2026.
- Longest answer
- 131,072 tokens
- The gateway's output ceiling per request. Measured Sep 3, 2026.
Read about TheoArca Velox
The fine print, up front
What our own model means here
01
Ours, end to end
Every model on this page runs on hardware OpenCharts operates and carries the Theo name because we stand behind it by name. Nothing you send to one of them goes to a third-party model provider, and each engine is versioned by us: it changes only in a deliberate release, never by drift.
02
Every number is dated
The figures on these pages are our own measurements on our own hardware: a render time, a capture duration, a review round. Each one says when it was measured and where the number came from. None of them is a benchmark score, and none is rounded into a marketing number.
03
Safety is built in, not bolted on
A model we run is a model we are accountable for. Our image engine screens the prompt before the render and classifies the image after it, and fails closed when a detector cannot answer. Every page states what is kept, for how long, and why.
On the benchmarks page
Measured by us, marked as such
Our text engines sit on the OpenCharts Index as first-party rows. Their scores are our own runs of the public tests and stay labelled internal until a third party publishes one. They never lead a spotlight.
TheoArca Magnus 5.1
Every benchmark reads TBD until a run is recorded.
No run recorded yetTheoArca Velox 5.1
Every benchmark reads TBD until a run is recorded.
No run recorded yetSnapshot 2026-09-08. The image and capture engines are not ranked there: the Index ranks text models on third-party published tests. How the Index works
Where to find them
Inside Theo today. Elsewhere as it becomes real.
Every model lives in Theo first. A public API and third-party evaluation are listed the day they exist and marked in progress or planned until then; nothing here links to a page that is not there yet.
- LiveInside Theo. In Theo Stealth today. Open
- PlannedPublic API. A metered endpoint with the same prompt screen and output classifier in front of it.
- PlannedThird-party arenas. Follows the public API: arenas rank publicly callable models.
- Rolling outInside Theo. The Real World mode in Theo Chat, where enabled.
- PlannedThird-party evaluation. Reconstruction quality is measured against held-out frames today; a public evaluation follows.
- Rolling outInside Theo. The composer's Theo Arca pick, an engine pin in Settings and the Code Canvas compliance tiers, where enabled.
- In progressPublic API. Served through the TheoVex gateway today; a public, metered listing on model marketplaces is in progress.
- PlannedThird-party evaluation. Follows the public listing. Until then the benchmarks page shows OpenCharts' own runs, labelled internal.
- Rolling outInside Theo. The composer's Theo Arca pick, an engine pin in Settings and the Code Canvas compliance tiers, where enabled.
- In progressPublic API. Served through the TheoVex gateway today; a public, metered listing on model marketplaces is in progress.
- PlannedThird-party evaluation. Follows the public listing. Until then the benchmarks page shows OpenCharts' own runs, labelled internal.
Questions
Questions? Answered
An engine OpenCharts runs on its own hardware, or through its own lab, and stands behind by name: the pipeline, the safety controls, the retention and the versioning are ours, and nothing you send to it goes to a third-party model provider. Each page says what the engine does, what it guarantees and what we measured, in plain words.
No. Theo is an orchestrator: it runs on frontier engines from several labs and shows which one handled each step. The models on this page are the ones we build or run ourselves, and each page says exactly where it appears: a Stealth engine pick, a chat mode, a composer pick or a Code Canvas tier.
On our own hardware, by us, and each one is dated with its source. A render time was timed on the hardware that serves the engine; a capture duration was timed end to end on the production path; a safety round is a human review of a labelled corpus. They are measurements, not benchmarks, and they are re-measured rather than extrapolated when something changes.
The text engines already do, as first-party rows whose scores are OpenCharts' own runs, labelled internal until a third party publishes a score. The image and capture engines are not ranked there: the OpenCharts Index ranks text models on third-party published tests, and there is no equivalent for an image or a capture engine yet.
Not yet. Every page carries a Where to find it section that says what is live, what is in progress and what is planned, and links only to what exists. When a channel goes live, it is linked there the same day.
No. The pages describe what an engine does, what it guarantees and what we measured, not how it is assembled. How our models are built is our own work and stays with us; what you can hold us to is written on every page: the safety controls, the retention, the versioning and the dated numbers.
No. A white-label tenant's AI is its own product, so this hub, like the benchmarks hub, is not served on tenant hosts.
Built here, measured here, shown here.
One of these engines is in Theo today, and every one says what it is. Theo picks the right engine for each step and always shows which one answered.